Extreme rainfall intensity inception nowcasting method based on deep learning
By combining multi-source data of stationary satellites, ground automatic stations and numerical modes, and using deep learning models for feature extraction and fusion, the accuracy and waste of computing resources of extreme rain strong first-generation forecasts are solved, and efficient extreme rain strong first-generation forecasts are achieved.
Patent Information
- Application Number
- CN202510610263.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-13
- Publication Date
- 2025-08-08
AI Technical Summary
The prior art is difficult to effectively combine multi-source heterogeneous data of stationary satellites, ground automatic stations and numerical forecasting modes to accurately predict extreme rain and waste of computing resources.
A deep learning model is used to combine multi-source data of stationary satellites, ground automatic stations and high-resolution regional numerical modes, and feature extraction and fusion are performed through the transformer model of the self-attention mechanism to identify the characteristics related to extreme rain and strong initial generation, and output the probability of occurrence.
The hit rate of the extreme rain strong newborn is improved, the false alarm rate is reduced, and the accuracy and efficiency of the forecast is improved.
Smart Images
Figure CN120447108A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of atmospheric science and remote sensing research, and also relates to the field of information technology processing, and more specifically to a method for the initial and nowcasting of extreme rainfall intensity based on deep learning. Background Art
[0002] According to the criteria for determining extreme heavy rainfall events, extreme rainfall intensity is defined as a low-probability heavy rainfall phenomenon in which precipitation in a specific area significantly exceeds the average for the local climate standard period within a specific time period. This phenomenon is statistically rare and is typically characterized by high destructiveness, suddenness, and high uncertainty. Accurate forecasts of extreme rainfall intensity are crucial for disaster prevention and mitigation.
[0003] Given the uncertainty of weather forecasts, predicting low-probability extreme precipitation events presents a significant challenge. Current extreme precipitation diagnostic schemes are primarily based on the characteristics of environmental physical variables predicted by numerical models. These schemes rely heavily on the accuracy of numerical model predictions for extreme precipitation. However, deterministic forecasts based on numerical model predictions struggle to accurately capture low-probability extreme precipitation events. Numerical model ensemble forecasts improve extreme precipitation forecasting capabilities by characterizing the uncertainty of extreme precipitation forecasts. However, this requires significant computational resources, which can be a significant waste of computational resources for low-probability extreme precipitation events.
[0004] The rapid development of a new generation of high-temporal and spatial-resolution geostationary satellites, dual-polarization weather radars, ground-based automatic observations, and high-resolution regional numerical models has brought new opportunities for extreme precipitation forecasting. However, previous methods for identifying and forecasting extreme precipitation based on multispectral signal observations from geostationary satellites have mostly used empirical thresholds to identify and predict the occurrence of extreme precipitation. This requires different thresholds for different regions, significantly limiting their applicability. Furthermore, previous methods have rarely considered the role of surface-based characteristic parameters derived from densely populated ground-based automatic observations in identifying and forecasting the onset of extreme precipitation. Summary of the Invention
[0005] The purpose of the present invention is to provide a method for the immediate forecast of extreme rainfall intensity based on deep learning, which combines multi-source heterogeneous data from geostationary satellites, ground automatic stations and numerical forecast models, and uses deep learning models to extract and fuse multiple features, so as to realize the immediate forecast of extreme rainfall intensity and improve the hit rate of the immediate forecast of extreme rainfall intensity.
[0006] To achieve the above objectives, the present invention provides a deep learning-based extreme rainfall intensity nowcasting method, comprising:
[0007] S1. Obtain observation data from geostationary satellites and automatic ground stations, as well as data predicted by high-resolution regional numerical models, and perform preprocessing.
[0008] S2. Based on observation data from geostationary satellites and ground-based automatic stations, a multi-threshold convection identification and tracking method is used to identify historical extreme rainfall events and determine the time and location of the onset of extreme rainfall.
[0009] S3. Extract geostationary satellite multispectral signals at several moments before the onset of extreme rainfall intensity from geostationary satellite observation data; calculate ground environmental parameters at the corresponding moments based on ground automatic station observation data; and calculate numerical model environmental parameters at the corresponding moments based on data predicted by a high-resolution regional numerical model;
[0010] S4. Using statistical methods such as principal component analysis and correlation analysis, sensitive factors closely related to the onset of extreme rainfall intensity are screened from the geostationary satellite multispectral signals, ground environmental parameters, and numerical model environmental parameters, and normalized to construct a data set.
[0011] S5. Based on the dataset established above, the self-attention mechanism transformer model is used to extract features related to extreme rainfall intensity, perform feature fusion, and output the probability of occurrence of the initial extreme rainfall intensity.
[0012] A further preferred technical solution of the present invention is that, in step S3, the geostationary satellite multispectral signal at a certain time before the onset of the extreme rainfall intensity is directly extracted from the geostationary satellite observation data; and the corresponding ground environmental parameters, including the dew point temperature, the one-hour temperature variation, and the one-hour pressure variation, are calculated based on the minute-level dense ground automatic station observation data;
[0013] Based on the data predicted by the high-resolution regional numerical model, the numerical model environmental parameters at the corresponding time are calculated, including the K index K, the convective effective potential energy CAPE, the convective inhibition energy CIN, the lifting index LI and the Schaffner index SI. The calculation formulas of each numerical model environmental parameter are:
[0014] K=(T 850 -T 500 )+T d850 -(TT d ) 700 ;
[0015]
[0016] LI=T e500 -T i500 ;
[0017] SI=T e500 -T i500 ;
[0018] Where T represents temperature, T d is the dew point temperature, T 850 is 850hPa temperature, T 500 is 500hPa temperature, T d850 is the dew point temperature at 850hPa, (TT d ) 700 Indicates the difference between 700hPa temperature and dew point temperature, T e500 The ambient temperature is 500hPa, T i500 It represents the temperature of the air mass when it rises along the dry adiabatic line, reaches the lifting condensation height, and then rises along the wet adiabatic line to 500hPa;
[0019] g is the gas constant, Z EL For the balance height, Z LFC is the free convection height, T vp is the virtual temperature associated with the air mass, T ve is the virtual temperature related to the environment, Z l is the initial rising height of the air mass, R d is the gas constant for dry air.
[0020] Preferably, the time resolution of geostationary satellite observation data is 10 minutes. When calculating ground environmental parameters, minute-level parameters will be aligned with geostationary satellite observations; for hour-level parameters, the time of the onset of extreme rainfall intensity will be used as the benchmark, and observation data from several moments before will be extracted to calculate the corresponding ground environmental parameters.
[0021] High-resolution regional numerical model forecasts are all hourly forecasts at the hour. Similarly, the numerical model forecasts for the hour several moments before the occurrence of extreme rainfall intensity are extracted based on the time when the extreme rainfall intensity first occurs. Time coding technology is used to take the time information of the corresponding moment of the numerical model environmental parameters as an additional constraint condition.
[0022] Preferably, the self-attention mechanism transformer model in step S5 dynamically adjusts the importance of features by calculating the correlation between input features, and automatically learns the dependency between features at different time and space scales. Its core formula is:
[0023]
[0024] Where Q is the query, K is the key, V is the value, d k is the dimension of the key vector.
[0025] As a preferred method, considering the characteristic differences between different observation data, different feature encoding branches are used to extract features related to extreme rainfall intensity from geostationary satellite multispectral signals, ground environmental parameters, and numerical model environmental parameters respectively. In each branch, a patch embedding module is used to cut different variables into slices of fixed size, and then position encoding is used to match each slice. The slices are input into the encoding module together with the corresponding slices to extract high-dimensional refined features of geostationary satellite multispectral signals, ground environmental parameters, and numerical model environmental parameters respectively.
[0026] Then, the extreme rainfall intensity-related features extracted from different branches are fused through the multi-head attention mechanism module. Finally, the occurrence probability of extreme rainfall intensity is output in the forecast module.
[0027] Preferably, the time resolution of the geostationary satellite multispectral signal is 10 minutes, the time resolution of the ground environmental parameters is 1 hour, and the time resolution of the numerical model environmental parameters is 1 hour. In each feature extraction module, an additional time coding unit will be introduced, and the time information of the corresponding observation will be used as an additional constraint.
[0028] Preferably, the method for preprocessing the observation data obtained from geostationary satellites and ground automatic stations, and the data predicted by high-resolution regional numerical models in step S1 includes checking the integrity and consistency of the data, and performing quality control on the data.
[0029] Preferably, in step S2, while using a multi-threshold convection identification and tracking method to identify historical extreme rainfall events, historical non-extreme rainfall events are also identified; in step S3, geostationary satellite multispectral signals, ground environmental parameters and numerical model environmental parameters at several moments before the onset of non-extreme rainfall are simultaneously obtained; in step S4, sensitive factors closely related to the onset of non-extreme rainfall are screened out and added to the data set.
[0030] Preferably, in step S4, a principal component analysis method or a Pearson correlation coefficient method is used to screen out sensitive factors closely related to the onset of extreme rainfall intensity from the geostationary satellite multispectral signals, ground environmental parameters and numerical model environmental parameters.
[0031] Beneficial effects: The present invention comprehensively considers the multi-spectral signals of geostationary satellites, ground environmental characteristic parameters, and high-temporal and spatial resolution regional numerical model environmental characteristic parameters, and combines the transformer deep learning model with the self-attention mechanism to improve the extraction and fusion capabilities of the incipient features related to extreme rainfall intensity, which helps to improve the hit rate of the incipient probability forecast of extreme rainfall intensity, reduce the false alarm rate of the forecast, and improve the forecast skills of the incipient probability forecast of extreme rainfall intensity.
[0032] The present invention processes multi-source heterogeneous data from geostationary satellites, ground automatic stations, etc. through multi-branch coding, thereby enhancing the feature extraction and fusion capabilities of geostationary satellite multispectral signals, ground environmental characteristic parameters, and high-resolution regional numerical model environmental characteristic parameters related to the onset of extreme rainfall intensity, further improving the hit rate of extreme rainfall intensity onset forecast, and reducing the false alarm rate of extreme rainfall intensity onset forecast, so as to more accurately forecast the onset of extreme rainfall intensity. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 Flowchart of a method for extreme rainfall intensity initial nowcasting based on deep learning in an embodiment of the present invention;
[0034] Figure 2 This is a data processing flow chart of the self-attention mechanism transformer model in an embodiment of the present invention. DETAILED DESCRIPTION
[0035] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the embodiments described are part of the embodiments of the present invention, not all of the embodiments, and they should not be understood as limitations on the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In the description of the present invention, it should be understood that the terms used are only for descriptive purposes and cannot be understood as indicating or implying relative importance.
[0036] Embodiment: The technical problem of the present invention is: a deep learning-based extreme rainfall intensity initial nowcasting method, combining multi-source heterogeneous data from geostationary satellites, ground automatic stations and numerical forecast models, and using a deep learning model to extract and fuse multiple features, to achieve the immediate forecast of the extreme rainfall intensity and improve the hit rate of the immediate forecast of the extreme rainfall intensity.
[0037] How to effectively fuse multi-source heterogeneous data to improve forecast accuracy is the key difficulty of the present invention. Specifically, the three types of data, namely geostationary satellite multispectral signals, ground environmental parameters and numerical model environmental parameters, have significant differences in spatiotemporal resolution, data structure and physical significance. Traditional methods find it difficult to fully utilize the complementary information in these heterogeneous data. In particular, when processing two-dimensional image data from geostationary satellites, how to retain its spatial structure information and organically fuse it with other one-dimensional time series data becomes a major challenge. In addition, the initial occurrence of extreme rainfall events has the characteristics of suddenness and localization. How to quickly identify and extract key features related to the event from massive data is also a difficulty. In terms of model design, it is necessary to consider how to reduce the computational complexity of the model while ensuring forecast accuracy to meet the real-time requirements of business operations. At the same time, due to the rarity of the initial occurrence of extreme rainfall events, how to solve the sample imbalance problem and improve the model's ability to identify low-probability events are also problems that need to be solved urgently.
[0038] Therefore, this embodiment provides a method for extreme rainfall intensity initial forecasting based on deep learning. The overall steps are as follows: Figure 1 As shown in the figure, first, based on multi-source observations from FY-4A / B geostationary satellites, ground automatic stations, and surface rain gauges, the corresponding multispectral signals of rapid cumulus cloud development and ground environmental characteristic parameters are extracted. At the same time, the relevant environmental characteristic parameters from the corresponding high-resolution regional numerical model are extracted to establish a set of initial extreme rainfall intensity data sets. Secondly, a transformer model with an attention mechanism and a multi-branch structure is used to extract and fuse features from the multispectral signals of heterogeneous data from multiple sources such as geostationary satellites, ground automatic stations, and surface rain gauges, ground environmental characteristic parameters, and environmental characteristic parameters from the high-resolution regional numerical model, to generate an initial extreme rainfall intensity probability forecast product.
[0039] The deep learning extreme rainfall intensity nowcasting method proposed in this embodiment is described in detail, and the steps include:
[0040] S1. Collect observation data from FY-4A / B geostationary satellites and ground automatic stations (including temperature, air pressure, relative humidity, wind direction, wind speed, precipitation, etc.), check the integrity and consistency of the data, and perform quality control.
[0041] S2. Based on the above observation data, a multi-threshold convection identification and tracking method is used to identify and track severe convective events. Then, the corresponding precipitation is matched according to the mask object at each moment of identification to determine severe convective events with precipitation exceeding 50 mm / h (defined as extreme rainfall intensity precipitation events) and convective events with precipitation less than 50 mm / h (defined as non-extreme rainfall intensity precipitation events). Then, the time and location when the precipitation of extreme rainfall intensity precipitation events first exceeds 50 mm / h are determined, that is, the onset of extreme rainfall intensity.
[0042] S3. Based on the extracted information on the onset of extreme rainfall intensity, the cloud masks of the identified and tracked extreme rainfall events are matched, and multispectral signals such as cloud top height and cloud phase are directly extracted from the FY-4A / B geostationary satellite observations 1-2 hours before the onset of extreme rainfall intensity. The corresponding ground environmental characteristic parameters are calculated based on intensive minute-level ground automatic station observations, including dew point temperature, 1-hour temperature change, and 1-hour pressure change. The dew point temperature is the directly observed variable, and the 1-hour temperature change and 1-hour pressure change are the rates of change of air temperature and air pressure with a time interval of 1 hour.
[0043] At the same time, the data predicted by the regional numerical model with high temporal and spatial resolution (time resolution of 1 hour and spatial resolution of 1 km) are used to extract the environmental characteristic parameters at the corresponding time, including the K index K, convective effective potential energy CAPE, convective inhibition energy CIN, lift index LI and Sha's index SI. The calculation formulas of the environmental parameters of each numerical model are:
[0044] K=(T 850 -T 500 )+T d850 -(TT d ) 700 ;
[0045]
[0046] LI=T e500 -T i500 ;
[0047] SI=T e500 -T i500 ;
[0048] Where T represents temperature, T d is the dew point temperature, T 850 is 850hPa temperature, T 500 is 500hPa temperature, T d850 is the dew point temperature at 850hPa, (TT d ) 700 Indicates the difference between 700hPa temperature and dew point temperature, T e500 The ambient temperature is 500hPa, T i500 It represents the temperature of the air mass when it rises along the dry adiabatic line, reaches the lifting condensation height, and then rises along the wet adiabatic line to 500hPa;
[0049] g is the gas constant, Z EL For the balance height, Z LFC is the free convection height, T vp is the virtual temperature associated with the air mass, T ve is the virtual temperature related to the environment, Z lis the initial rising height of the air mass, R d is the gas constant for dry air.
[0050] The calculation formulas for the lift index LI and the Sha's index SI are the same, the main difference being the starting lift height.
[0051] The acquired geostationary satellite multispectral signals, ground environmental parameters and numerical model environmental parameters are shown in Table 1.
[0052] Table 1 Geostationary satellite multispectral signals, ground environmental parameters and numerical model environmental parameters
[0053]
[0054] In addition, geostationary satellite multispectral signals, ground environmental parameters, and numerical model environmental parameters corresponding to non-extreme rainfall events are extracted. It is important to note that when matching regional numerical model forecasts, the nearest hourly forecast 1-2 hours before the onset of the corresponding extreme rainfall event and non-extreme rainfall event is matched. In addition to extracting the corresponding environmental characteristic parameters, the corresponding forecast time is recorded.
[0055] Since the time resolution of geostationary satellite observations is 10 minutes, when calculating ground characteristic parameters, minute-level characteristic parameters will be aligned with geostationary satellite observations. For hourly characteristic parameters, the time of occurrence of extreme rainfall intensity and non-extreme rainfall intensity events will be used as a benchmark to extract observations 1-2 hours ago and calculate the corresponding characteristic parameters, such as water vapor at different heights in the troposphere, cloud top heights at different altitudes, cloud top phase, K index, convective effective potential, etc. Considering that high-resolution regional numerical models are hourly forecasts at the hour, in order to align with the above observations in time, the time of occurrence of extreme rainfall intensity and non-extreme rainfall intensity events will also be used as a benchmark to extract the numerical model forecast at the hour 1-2 hours ago. In order to avoid introducing additional errors, in this embodiment, the numerical model environmental characteristic parameters are not interpolated to the same time resolution as the above observations using an interpolation method. Instead, a time encoding technique is used to use the time information of the corresponding moment of the regional numerical model environmental characteristic parameters as an additional constraint.
[0056] S4. Based on the extracted geostationary satellite multispectral signals of extreme rainfall intensity and non-extreme rainfall intensity, ground environmental parameters, and numerical model environmental parameters, principal component analysis or Pearson correlation coefficient or other correlation statistical methods are used to screen sensitive factors that are closely related to the initial occurrence of extreme rainfall intensity; then, based on the screened sensitive factors, the corresponding means and standard deviations are calculated to normalize the above input factors;
[0057] The normalized dataset is divided into training, validation, and test sets in chronological order, with April to September 2020-2022 as the training set, April to September 2023 as the validation set, and April to September 2024 as the test set. The input factors are geostationary satellite multispectral signals, ground environmental characteristic parameters, and regional numerical model environmental characteristic parameters related to the onset of extreme rainfall intensity, and the corresponding labels are extreme rainfall intensity onset and non-extreme rainfall intensity events. Note: When constructing the dataset, it is necessary to ensure that different sub-datasets are independent of each other to avoid information leakage. At the same time, in addition to the onset of extreme rainfall intensity events, each sub-dataset should also include non-extreme rainfall intensity onset events (i.e., precipitation events less than 20 mm / h), as well as events at different times before and after the onset of extreme rainfall intensity, so as to ensure the diversity of samples in the dataset and enhance the generalization ability of the model.
[0058] S5. Based on the above-established dataset, Figure 2 As shown in the figure, the self-attention mechanism transformer model is used to extract features related to extreme rainfall intensity, and feature fusion is performed to output the probability of the occurrence of extreme rainfall intensity.
[0059] The self-attention mechanism transformer model dynamically adjusts the importance of features by calculating the correlation (weight) between input features, and automatically learns the dependencies between features at different time and space scales. Its core formula is:
[0060]
[0061] Among them, Q is the query, K is the key, V is the value, d k is the dimension of the key vector.
[0062] Taking into account the characteristic differences between different observation data, different feature encoding branches are used to extract features related to extreme rainfall intensity from geostationary satellite multispectral signals, ground environmental parameters and numerical model environmental parameters, respectively. In each branch, a patch embedding module is used to divide different variables into slices of fixed size. Then, position encoding is used to match each slice and input it into the encoding module together with the corresponding slice to extract high-dimensional and refined features of geostationary satellite multispectral signals, ground environmental parameters and numerical model environmental parameters, respectively.
[0063] It should be noted that since the time resolution of geostationary satellite multispectral signals and ground environmental characteristic parameters is inconsistent with the time resolution of regional numerical model environmental characteristic parameters, the time resolution of geostationary satellite observations is 10 minutes, the time resolution of ground environmental characteristic parameters is 1 hour, and the time resolution of numerical model environmental characteristic parameters is 1 hour. In each feature extraction module, an additional time coding unit will be introduced, and the time information of the corresponding observation will be used as an additional constraint, so as to realize the alignment of multi-source data with inconsistent time in high-dimensional features and alleviate the additional error introduced by direct interpolation alignment time.
[0064] Then, the extreme rainfall intensity-related features extracted from different branches are fused through the multi-head attention mechanism module. Finally, the occurrence probability of extreme rainfall intensity is output in the forecast module.
[0065] The method for the near-occurrence forecast of extreme rainfall intensity provided in this embodiment obtains environmental characteristic parameters of geostationary satellite multispectral signals, ground automatic stations and high-resolution regional numerical models, adopts a multi-threshold convection identification and tracking method to identify severe convective events, and determines whether it is an extreme rainfall intensity first-occurrence event based on precipitation. Multi-source data is processed using a patch embedding module and a time coding unit, and a transformer model based on a self-attention mechanism is introduced to extract high-dimensional features. Feature fusion is performed through a multi-head attention mechanism, and finally the probability of the occurrence of extreme rainfall intensity first-occurrence is output. The method of this embodiment enhances the feature extraction and fusion capabilities of multi-source data, improves the hit rate of extreme rainfall intensity first-occurrence near-occurrence forecasts, reduces the false alarm rate, and provides an effective method for more accurately forecasting the onset of extreme rainfall intensity.
[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.
Claims
1. A deep learning-based extreme rainfall intensity nowcasting method, characterized by: include: S1. Obtain observation data from geostationary satellites and ground automatic stations, as well as data predicted by high-resolution regional numerical models, and perform preprocessing; S2. Based on observation data from geostationary satellites and ground-based automatic stations, a multi-threshold convection identification and tracking method is used to identify historical extreme rainfall events and determine the time and location of the onset of extreme rainfall. S3. Extract geostationary satellite multispectral signals at several moments before the onset of extreme rainfall intensity from geostationary satellite observation data; calculate ground environmental parameters at the corresponding moments based on ground automatic station observation data; and calculate numerical model environmental parameters at the corresponding moments based on data predicted by a high-resolution regional numerical model; S4. Using principal component analysis or correlation analysis, screen out sensitive factors closely related to the onset of extreme rainfall intensity from the geostationary satellite multispectral signal, ground environmental parameters, and numerical model environmental parameters, perform normalization, and construct a data set; S5. Based on the dataset established above, the self-attention mechanism transformer model is used to extract features related to extreme rainfall intensity, perform feature fusion, and output the probability of occurrence of extreme rainfall intensity.
2. The deep learning-based extreme rainfall intensity nowcasting method according to claim 1, characterized in that: In step S3, the geostationary satellite multispectral signal at several moments before the onset of extreme rainfall intensity is directly extracted from the geostationary satellite observation data; then, the corresponding ground environmental parameters, including dew point temperature, 1-hour temperature variation, and 1-hour pressure variation, are calculated based on the minute-level dense ground automatic station observation data; Based on the data predicted by the high-resolution regional numerical model, the numerical model environmental parameters at the corresponding time are calculated, including the K index K, the convective effective potential energy CAPE, the convective inhibition energy CIN, the lifting index LI and the Schaffner index SI. The calculation formulas of each numerical model environmental parameter are: K=(T 850 ―T 500 )+T d850 ―(T―T d ) 700 ; LI=T e500 ―T i500 ; SI=T e500 ―T i500 ; Where T represents temperature, T d is the dew point temperature, T 850 is 850hPa temperature, T 500 is 500hPa temperature, T d850 is the dew point temperature at 850hPa, (T-T d ) 700 Indicates the difference between 700hPa temperature and dew point temperature, T e500 The ambient temperature is 500hPa, T i500 It represents the temperature of the air mass when it rises along the dry adiabatic line, reaches the lifting condensation height, and then rises along the wet adiabatic line to 500hPa; g is the gas constant, Z EL For the balance height, Z LFC is the free convection height, T vp is the virtual temperature associated with the air mass, T ve is the virtual temperature related to the environment, Z l is the initial rising height of the air mass, R d is the gas constant for dry air.
3. The deep learning-based extreme rainfall intensity nowcasting method according to claim 2, characterized in that: The time resolution of geostationary satellite observation data is 10 minutes. When calculating ground environmental parameters, minute-level parameters will be aligned with geostationary satellite observations. For hourly parameters, the time of the onset of extreme rainfall intensity will be used as the benchmark, and observation data from several moments before will be extracted to calculate the corresponding ground environmental parameters. High-resolution regional numerical model forecasts are all hourly forecasts at the hour, and the numerical model forecasts for the hour several moments before the onset of extreme rainfall intensity are also extracted; Time coding technology is used to take the time information of the numerical model environmental parameters at the corresponding moment as an additional constraint condition.
4. The deep learning-based extreme rainfall intensity nowcasting method according to claim 1, characterized in that: In step S5, the self-attention mechanism transformer model dynamically adjusts the importance of features by calculating the correlation between input features, and automatically learns the dependencies between features at different time and space scales. Its core formula is: Where Q is the query, K is the key, V is the value, d k is the dimension of the key vector.
5. The deep learning-based extreme rainfall intensity nowcasting method according to claim 4, characterized in that: Considering the differences in characteristics between different observational data, different feature encoding branches are used to extract features related to extreme rainfall intensity from geostationary satellite multispectral signals, ground environmental parameters, and numerical model environmental parameters. In each branch, a patch embedding module is used to divide different variables into slices of fixed size. Then, position encoding is used to match each slice and input it into the encoding module together with the corresponding slice to extract high-dimensional and refined features of geostationary satellite multispectral signals, ground environmental parameters, and numerical model environmental parameters. Then, the extreme rainfall intensity-related features extracted from different branches are fused through the multi-head attention mechanism module. Finally, the occurrence probability of extreme rainfall intensity is output in the forecast module.
6. The deep learning-based extreme rainfall intensity nowcasting method according to claim 5, characterized in that: The time resolution of geostationary satellite multispectral signals is 10 minutes, the time resolution of ground environmental parameters is 1 hour, and the time resolution of numerical model environmental parameters is 1 hour. In each feature extraction module, an additional time coding unit will be introduced, and the time information of the corresponding observation will be used as an additional constraint.
7. The deep learning-based extreme rainfall intensity nowcasting method according to claim 1, characterized in that: The method for preprocessing the observation data obtained from the geostationary satellite and the ground automatic station, as well as the data predicted by the high-resolution regional numerical model in step S1, includes checking the integrity and consistency of the data, and performing quality control on the data.
8. The deep learning-based extreme rainfall intensity nowcasting method according to claim 1, characterized in that: In step S2, while using a multi-threshold convection identification and tracking method to identify historical extreme rainfall events, historical non-extreme rainfall events are also identified; in step S3, geostationary satellite multispectral signals, ground environmental parameters, and numerical model environmental parameters at several moments before the onset of non-extreme rainfall are simultaneously acquired; in step S4, sensitive factors closely related to the onset of non-extreme rainfall are screened and added to the dataset.
9. The deep learning-based extreme rainfall intensity nowcasting method according to claim 1, characterized in that: In step S4, a principal component analysis method or a Pearson correlation coefficient method is used to screen out sensitive factors closely related to the onset of extreme rainfall intensity from the geostationary satellite multispectral signals, ground environmental parameters and numerical model environmental parameters.